Highlight extraction

What this does

Someone hands you a source — an article, a book chapter, a meeting transcript, a saved thread — and wants the standout passages captured as a list of highlights they can skim later. This skill produces that list: each entry is a passage copied exactly from the source, paired with a location tag saying where in the source it sits. Nothing is reworded, nothing is compressed into your own summary, and nothing is added that the source doesn't contain.

Think of it as an honest highlighter. The output is not analysis and not a summary — it's the selected words of the source, pulled out and labeled, so a reader can trust that every line is genuinely there and can find it again.

It targets the default failure of a bare model asked to "grab the key bits": it tightens the quote (drops filler, fixes grammar, swaps a word), fuses two separate sentences into one seamless-looking quote, slides into summary and paraphrases the point instead of copying it, and — worst — invents a location tag (p. 4, line 30) the source never shows. The result looks like clean notes but isn't faithful: a "highlight" may be something the author never wrote, tagged to a place that doesn't exist.

When to use / when NOT to

Use when:

• You're handed a source and asked to capture, pull, collect, or list the notable / standout /

memorable passages as quotes.

• The reader wants to review the actual words later and jump back to where each came from.
• Fidelity matters — research reading, book notes, evidence capture, transcript review.

Do NOT use when:

• The ask is a summary in your own words that replaces the source — that's summarization, and it

is supposed to be non-verbatim.

• The ask is to mark up the source in place (bold the load-bearing sentences, keep the full text)

— that's progressive summarization, a different shape.

• The ask is to pull fixed structured fields (invoice number, date, total) into JSON — that's

document field extraction, not passage highlighting.

• The ask is to translate the passages — the output is inherently not verbatim.

The contract (this is the lift)

1. Verbatim, character-exact. Each highlight is the source's own words, reproduced letter for

letter. Do not fix grammar, drop filler, modernize a word, change tense, or "tighten" for readability. If the source says Me and him wasn't ready, that is the highlight — as written.

1. No invention. Every highlight must appear in the source. Never add a sentence, a clause, or a

word the source doesn't contain, and never wrap the writer's own gloss as if it were a quote.

1. One passage per highlight — don't fuse. If two striking sentences sit apart in the source,

they are two highlights, not one. Only join spans inside a single highlight when you cut the middle, and then mark the cut with an ellipsis …. Never stitch non-adjacent text into one seamless-looking quote.

1. A location tag on every highlight, taken from the source's own signals. Tag each highlight

with where it sits, using only a locator the source actually exposes (see next section). If the source gives no location signal at all, tag by its position in the source (e.g. paragraph 3) — never print a page number, line number, or timestamp the source does not show.

1. Select, don't dump. Pull the notable passages — the vivid, surprising, load-bearing lines —

not every sentence. Boilerplate, throat-clearing, and routine connective text stay out.

1. Highlights only — no commentary inside the quote. The output is the passages and their tags.

If you must add a note about why a line matters, keep it in a separate labeled field, plainly outside the quoted words — never blended into the passage.

1. In source order. List the highlights in the order they appear in the source, so the tags run

forward.

Choosing the location tag — use what the source gives you

Read what location signal the source actually carries, and tag with that. In rough priority:

| The source shows… | Tag with | Example tag | |---|---|---| | Page markers / p. breaks | the page | p. 12 | | A transcript with timestamps | the timestamp | 00:14:32 | | Section / chapter headings | the heading | under "Methods" | | Numbered lines | the line number | line 47 | | None of the above | ordinal position | paragraph 3 |

The rule behind the table: the tag must be verifiable against the source. If the reader opens the source, the tag has to point them to the right spot. A page number on a source with no pages, or a timestamp on a plain text file, is a fabricated tag — as dishonest as a fabricated quote. When unsure, fall back to paragraph position, which any prose source supports.

Trimming inside one highlight

Sometimes a single sentence is worth capturing but has a droppable middle. You may cut it within one highlight if you mark the cut:

• Remove interior words with an ellipsis: "The rollout … failed on the third attempt."
• Never remove words silently — an unmarked highlight must match the source exactly.
• Do not use an ellipsis to glue two separate sentences together; that's fusing (rule 3). The

ellipsis marks a cut inside one continuous passage, not a bridge between two.

Worked examples

Assume this source for examples 1–4 (a plain text note, no page numbers):

[¶1] The committee met for six hours on Tuesday.
[¶2] "We were," the chair admitted, "completely unprepared for the scale of it."
[¶3] Routine budget items were approved without discussion.
[¶4] By the end, one member called the process "a slow-motion accident nobody stopped."

1 — Verbatim, tagged by paragraph

GOOD:

• "We were … completely unprepared for the scale of it." — paragraph 2
• "a slow-motion accident nobody stopped" — paragraph 4

BAD:

• "The chair admitted the committee was unprepared for the scale of the problem." — paragraph 2

The BAD version reworded the line (summary, not a highlight) and it's tagged as if it were a copied quote.

2 — Don't fuse separate passages

BAD: "We were completely unprepared … a slow-motion accident nobody stopped." — paragraphs 2–4

Those are two different sentences from two different paragraphs stitched into one quote. Keep them separate, each with its own tag.

3 — Don't invent a location

BAD: "a slow-motion accident nobody stopped" — p. 3, line 18

The source has no pages and no line numbers. Tag it paragraph 4 — the locator the source actually supports.

4 — Leave the boilerplate out

The line Routine budget items were approved without discussion. is unremarkable connective text — it should not appear as a highlight. Select the vivid, load-bearing lines, not every sentence.

5 — Keep colloquial wording exactly

Source (transcript, timestamped): [00:09:12] Witness: "Me and my crew, we wasn't never told nothing." GOOD: "Me and my crew, we wasn't never told nothing." — 00:09:12 BAD: "My crew and I were never told anything." — 00:09:12 — the grammar was corrected; a highlight is copied, not cleaned.

Common mistakes (what a bare model does)

1. Tightening the quote — dropping filler or fixing grammar so the highlight no longer matches.
2. Sliding into summary — restating the point in your own words instead of copying the passage.
3. Fusing — merging two non-adjacent sentences into one seamless-looking quote.
4. Fabricating a tag — printing a page/line/timestamp the source doesn't expose.
5. Missing tags — a bare list of quotes with no "where from."
6. Dumping everything — highlighting routine text instead of selecting the notable lines.
7. Blending commentary — mixing your gloss into the quoted words rather than a separate note.
8. Reordering — listing highlights out of the order they appear in the source.

Edge cases

• No location signal at all: tag by paragraph position. Never upgrade to a page/line number to

look precise.

• A whole paragraph is the highlight: capture it verbatim and tag the paragraph; don't paraphrase

it down.

• Passage spans a page break (source has pages): tag the page it starts on, or pp. 4–5; keep the

text exact across the break.

• Source has a typo: reproduce it as-is (you may append [sic]); do not silently fix it inside a

highlight.

• You want to explain why a line matters: put it in a separate note: line beside the highlight,

never inside the quoted words.

• Nothing notable in a stretch of the source: skip it. Coverage is not the goal; faithful

selection is.

Quick reference

| Situation | Do | Don't | |---|---|---| | A standout line | copy it word-for-word + tag it | reword or tighten it | | Two separate striking lines | two highlights, two tags | fuse into one quote | | Cutting a droppable middle | ellipsis inside the one passage | drop words silently | | Source has no pages/lines | tag by paragraph position | invent p. 4 / line 30 | | Routine / boilerplate text | leave it out | highlight every sentence | | Your reason it matters | separate labeled note | blend it into the quote | | Order of the list | source order, tags run forward | reorder by importance |